Efficient Approximation of Deep ReLU Networks for Functions on Low Dimensional Manifolds